Ablate, Variate, and Contemplate: Visual Analytics for Discovering Neural Architectures.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning model performance relies heavily on precise configuration of layers and parameters.
• Current methods for model configuration lack systematic guidelines, leading to tedious manual programming or expensive automated approaches.

Purpose of the Study:

• To introduce REMAP, a visual analytics tool designed for rapid discovery of deep learning models.
• To enable model builders to efficiently explore and experiment with neural network architectures.

Main Methods:

• REMAP facilitates exploration of the deep learning parameter space via global inspection and local experimentation.
• Users can identify architecture clusters, perform ablation/variation experiments, and handcraft models using a graphical interface.
• The tool's design was informed by a study with four deep learning model builders.

Main Results:

• REMAP allows users to quickly build deep learning models without manual programming.
• The tool enables efficient discovery of performant neural network architectures.
• Demonstrated efficiency through a use case involving visual exploration and semi-automated searches.

Conclusions:

• REMAP significantly accelerates the process of deep learning model development.
• The visual analytics approach empowers users to efficiently discover and refine neural network architectures.
• REMAP addresses the need for systematic and efficient deep learning model configuration.